Academia.eduAcademia.edu

Figure 3 – uploaded by Christos K . Georgiadis

See full PDFdownloadDownload figure
The proposed method although it improves the accuracy of recommendations, is unable to provide recommendations to other users that do not have at least a number of co-rated items and a certain PCC similarity value. For this reason, a hybrid approach that can switch to PCC only if enough ratings are not available for the multi-level approach to provide recommendations. Algorithm 1 provides the hybrid approach.

Figure 3 The proposed method although it improves the accuracy of recommendations, is unable to provide recommendations to other users that do not have at least a number of co-rated items and a certain PCC similarity value. For this reason, a hybrid approach that can switch to PCC only if enough ratings are not available for the multi-level approach to provide recommendations. Algorithm 1 provides the hybrid approach.

Related Figures (17)

When recommendations need to be generated for a user, then the ratings are loaded into memory and a similarity function is used. The main par is how to estimate the similarity value between two users. This is called neighborhood identification and the job of the similarity function is to firstly identify a pre- specified of k nearest neighbors according to their similarity value. In present recommendations systems the value of k can vary from a few, possi bly 2 to 5, to as many as possible with the number ranging from 10 to 20 to 30 and so on up to hundreds of neighbors. A high number of neighbors doesn’t necessarily mean that the accuracy of the recommend ations will be high though.
When recommendations need to be generated for a user, then the ratings are loaded into memory and a similarity function is used. The main par is how to estimate the similarity value between two users. This is called neighborhood identification and the job of the similarity function is to firstly identify a pre- specified of k nearest neighbors according to their similarity value. In present recommendations systems the value of k can vary from a few, possi bly 2 to 5, to as many as possible with the number ranging from 10 to 20 to 30 and so on up to hundreds of neighbors. A high number of neighbors doesn’t necessarily mean that the accuracy of the recommend ations will be high though.
While other methods use the method of absolute ratings either by adding weights, or by manipulatin different variables to achieve a better result, our approach is based on a multi-level division. Furthermor« we argue that every user has a rating expression tha should not be punished that way. Also, we agre with PCC based similarity approaches that consider higher the similarity when the values of the co-rate items are as close as possible. Our proposed method from it and the number of co-rated items. In order to which is defined in equation 5. In this equation, T s heavily relies on PCC, the similarity value returne achieve this we first introduce a similarity functio ands for the total number of co-rated items, x is positive real number such as x € R, y is a positive real number such as y € R. In our proposed method w argue that the by dividing the algorithm in multip improved. To show the effectiveness of our method in this we made experiments using four levels. In thi context, at the first four steps the number of co-rated items is checked and if it is more than the pre specified thresholds (t/, 12, t3, t4), then it can proceed to the next step to check the similarity valu derived for the two users from PCC. For users that a sufficient number of co-rated items exist and th PCC similarity value is greater than a pre-specified threshold (y) then a list of recommendations | returned. Otherwise, for users that there is not available a sufficient number of co-rated items, zero valu is returned. Finally, t/, (2, 13 and ¢4 are natural numbers that represent the constraints put on the numbe of co-rated items for each level (t1 € N,t2 EN,t3 € e levels the accuracy of the recommendations 1 N,t4EN / t4>t3>12>11)
While other methods use the method of absolute ratings either by adding weights, or by manipulatin different variables to achieve a better result, our approach is based on a multi-level division. Furthermor« we argue that every user has a rating expression tha should not be punished that way. Also, we agre with PCC based similarity approaches that consider higher the similarity when the values of the co-rate items are as close as possible. Our proposed method from it and the number of co-rated items. In order to which is defined in equation 5. In this equation, T s heavily relies on PCC, the similarity value returne achieve this we first introduce a similarity functio ands for the total number of co-rated items, x is positive real number such as x € R, y is a positive real number such as y € R. In our proposed method w argue that the by dividing the algorithm in multip improved. To show the effectiveness of our method in this we made experiments using four levels. In thi context, at the first four steps the number of co-rated items is checked and if it is more than the pre specified thresholds (t/, 12, t3, t4), then it can proceed to the next step to check the similarity valu derived for the two users from PCC. For users that a sufficient number of co-rated items exist and th PCC similarity value is greater than a pre-specified threshold (y) then a list of recommendations | returned. Otherwise, for users that there is not available a sufficient number of co-rated items, zero valu is returned. Finally, t/, (2, 13 and ¢4 are natural numbers that represent the constraints put on the numbe of co-rated items for each level (t1 € N,t2 EN,t3 € e levels the accuracy of the recommendations 1 N,t4EN / t4>t3>12>11)
The MAE results for the MovieLens 100k dataset are shown in figure 1. It is shown that when the neighborhood is small, up to 20 users, our proposed method outperforms all the other recommendation methods. When the number of neighbors is getting higher, i.e 40 or more, we can see that our proposed method becomes less effective.
The MAE results for the MovieLens 100k dataset are shown in figure 1. It is shown that when the neighborhood is small, up to 20 users, our proposed method outperforms all the other recommendation methods. When the number of neighbors is getting higher, i.e 40 or more, we can see that our proposed method becomes less effective.
The MAE results for the MovieLens 1m dataset are shown in figure 2. It is shown that our proposed method ourperforms the other methods. Besides that, we can see that as the number of neighbors is getting higher our proposed method becomes more effective. We can see that the larger the neighbrhood grows the results become better for all methods. This is due to the fact that a larger number of users is available to his dataset compared to the 100k one.
The MAE results for the MovieLens 1m dataset are shown in figure 2. It is shown that our proposed method ourperforms the other methods. Besides that, we can see that as the number of neighbors is getting higher our proposed method becomes more effective. We can see that the larger the neighbrhood grows the results become better for all methods. This is due to the fact that a larger number of users is available to his dataset compared to the 100k one.
Figure 3. MAE results for the Jester dataset
Figure 3. MAE results for the Jester dataset
Figure 4. MAE results for the Epinions dataset The MAE results for the Jester dataset are shown in figure 3. It is shown that our proposed method ourperforms the other methods. However in all cases (although our method achieves a lower MAE) there aren’t any significant changes when the size of the neighborhood grows. It is also remarkable that all the alternative recommendation methods have a noteworthy difference from PCC. Moreover, the results are almost identical when the neighborhood grows, since the number of users in the dataset is very small.
Figure 4. MAE results for the Epinions dataset The MAE results for the Jester dataset are shown in figure 3. It is shown that our proposed method ourperforms the other methods. However in all cases (although our method achieves a lower MAE) there aren’t any significant changes when the size of the neighborhood grows. It is also remarkable that all the alternative recommendation methods have a noteworthy difference from PCC. Moreover, the results are almost identical when the neighborhood grows, since the number of users in the dataset is very small.
Figure 5. MAE results for the MovieTweetings dataset The MAE results for the MovieTweetings dataset are shown in figure 5. It is shown that our proposed method ourperforms the other methods. Besides that, we can see that as the number of neighbors is getting higher our proposed method becomes more effective. The MAE results for the Epinions dataset are shown in figure 4. It is shown that our proposed method ourperforms the other methods. However in all cases there aren’t any significant changes between our method and the others when the size of the neighborhood grows.
Figure 5. MAE results for the MovieTweetings dataset The MAE results for the MovieTweetings dataset are shown in figure 5. It is shown that our proposed method ourperforms the other methods. Besides that, we can see that as the number of neighbors is getting higher our proposed method becomes more effective. The MAE results for the Epinions dataset are shown in figure 4. It is shown that our proposed method ourperforms the other methods. However in all cases there aren’t any significant changes between our method and the others when the size of the neighborhood grows.
In figure 10 we can see the results obtained by using the Precision and Recall metrics using the MovieTweetings dataset. Sub figure 10,a represents the Precision results for 5 neighbors and 5 recommendations. Sub figure 10,b represents the Recall results for 5 neighbors and 5 recommendations. Sub figure 10,c represents the Precision results for 10 neighbors and 10 recommendations. Sub figure 10,d represents the Recall results for 10 neighbors and 10 recommendations. It is shown that our proposed method produces results of better quality in all cases.
In figure 10 we can see the results obtained by using the Precision and Recall metrics using the MovieTweetings dataset. Sub figure 10,a represents the Precision results for 5 neighbors and 5 recommendations. Sub figure 10,b represents the Recall results for 5 neighbors and 5 recommendations. Sub figure 10,c represents the Precision results for 10 neighbors and 10 recommendations. Sub figure 10,d represents the Recall results for 10 neighbors and 10 recommendations. It is shown that our proposed method produces results of better quality in all cases.
Related topics:
Computer ScienceUsabilityRecommender SystemsCollaborative FilteringRecommender System
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved